Notes from the founder · 13 October 2026 · 5 min read
AI made building cheap. It did not make learning cheap.
By James Warder
- Founder notes
- AI coding
- Focus
The month I counted
In July 2026 we were a small team building more products than I like to admit. AI coding agents made that possible. Features that used to take a sprint took an afternoon. When a competitor shipped something, we could match it within the week. We wrote a rule for it and called it the Superset Doctrine: build what nobody has first, then absorb everything the incumbents have as the baseline.
It felt like leverage, and for building it was. Then I tried to count something different. I did not count how much we shipped. I counted how many times reality had answered us. A loop, the way I mean it, has four steps: build something, let reality respond, learn from the response, change what you build. A user churning is a loop. A buyer saying no is a loop. A grant committee replying in three months is technically a loop, but a slow one.
Across everything we were building, we were getting about one or two real loops a month.
That number explained a lot. Our products were not bad. Most of them simply had no fast source of feedback: no users yet, or users we could only reach through a committee. A specification never pushes back. We could improve a document forever and learn nothing.
Two currencies
The mistake I had made is common now, so it is worth stating plainly.
| Did AI make it cheaper? | What bounds it | |
|---|---|---|
| Breadth: how much we can build | Yes, dramatically | Engineering hours, which we no longer lack |
| Depth: how many times we learn something real about one thing | No. AI speeds up the build step only | How fast reality answers, and the founder's attention |
Breadth is bought with AI. Depth is bought with attention. When building gets cheap, the temptation is to spend the cheap currency everywhere, because it feels like progress. But a product with a thousand loops behind it beats a product with a thousand features. And a small team cannot run a thousand loops on six things at once, because the slow steps happen outside the team.
The problems were not build failures
The other thing I noticed that month: none of our worst mistakes were engineering mistakes. They were coherence mistakes.
- Two documents written for the same reader that disagreed with each other.
- A description of one of our own contracts written from a design document instead of from the deployed code. It was wrong in three places, and wrong in a different way each time someone else rewrote it.
- An old repository, whose name matched the product, read as if it were the live system.
When we listed seven errors like these, five had the same cause: describing our own system from a document instead of from what was deployed. Since then, any claim about our own system needs a contract address, a transaction hash or a file and line behind it before it leaves the building. A PRD is not evidence. My memory is not evidence either.
AI makes this worse before it makes it better. It writes a confident paragraph about a system in seconds, and that paragraph reads exactly as well whether it came from the code or from a stale document.
What we changed
One product. We chose one product and gave it everything. The real argument for that was not focus as a virtue. It was arithmetic: if feedback loops are the scarce thing, every loop the company can generate should point at the same edge.
An outside date on every next step. Our clearest evidence came from our own product. Its specification was written on 8 September 2026 with no date attached. On 9 September a hackathon gave it one. The first version shipped on 13 September. Nothing else changed. Our product moves when an outside deadline is attached to it, and does not move otherwise. So the next increment gets a date before it gets a plan.
Parked means zero hours. We used to resolve hard choices by keeping every option alive, which in practice means giving each one 30% and none of them enough. Now every open bet is marked Active or Parked. Parked is not deleted. It stays fully written down and it gets no time this month. Keeping an option in the document is free. Keeping it in the week is not.
Every bet carries four questions: What would tell us we are wrong? How fast would it tell us? How many loops a month does it get? Is it Active or Parked this month? If the honest answer to the second is "a committee in November", the bet has no loop, and we stop pretending it does.
What I would tell another small team
Use AI for breadth. It is wonderful at parity work: the features you need so you do not lose. Then spend your own attention on the one thing nobody else has, and buy loops on it. Users, design partners, a public demo with a date, a buyer who might say no. Anything that answers faster than a committee.
Building is cheap now. Being told you are wrong is still expensive, and it is still the only thing that makes you better.
James Warder is the founder and CTO of Morca Labs.